PaddleOCR ONNX mirror

Byte-for-byte copies of the official PaddleOCR ONNX inference archives, mirrored because the upstream host is slow and highly variable outside China: the same file has measured 44 KB/s and 640 KB/s ten minutes apart, and a 21 MB model has taken over eight minutes to fetch.

Nothing here is modified. Each archive is the upstream .tar unchanged, and the SHA-256 below is the upstream file's own — the same value the consumer verifies after download, so a mismatch means this mirror is wrong, not that the model differs.

file bytes sha256
PP-OCRv5_mobile_det_onnx_infer.tar 4843520 781056046c9ed77a15c94681605db6a0f62317c2e9cce6931c71da2478d4bc30
PP-OCRv5_mobile_rec_onnx_infer.tar 16701440 f7e792bc836f36e7ef895ad47c426d75b0b75b1650caa6d63fe9418441ffba8c
korean_PP-OCRv5_mobile_rec_onnx_infer.tar 13537280 568ed8b43a260adc9f484d92105e425ea8cddf8ce16940c177bc12864cfb0eb0
PP-OCRv6_small_det_onnx_infer.tar 9891840 d218f6fbf0f1c23d2161bd6ac7f5eaa6104fa89955c09290497e31008e2618e4
PP-OCRv6_small_rec_onnx_infer.tar 21319680 d267ab077a44a0eedb1ea8f8c542d263f211de8e9d7a029bf9fcfff7e5a88fb1

Upstream

All five come from https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/, published by the PaddlePaddle team as part of PaddleOCR / PaddleX.

To re-verify any file against upstream:

curl -sSL -o f.tar "https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/<name>.tar"
shasum -a 256 f.tar

Who uses this

MetelScan's optional on-device OCR pipeline. It pins each file by revision and checks the SHA-256 after download, so this mirror cannot substitute different weights without the install failing.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support